Nvidia is in discussions to invest in Mercor, the AI data-labeling startup that supplies the specialized human-expert data Nvidia uses to train its own open-source Nemotron models, as part of a funding round that would value Mercor at 20 billion dollars, according to The Information. That figure would double Mercor’s valuation from just 10 billion dollars in October, less than a year ago, and it would make Nvidia a financial backer of a company it already pays tens of millions of dollars per quarter.
What Mercor actually does, and why Nvidia needs it
Mercor connects companies with human domain experts who label and generate the specialized training data modern AI models increasingly depend on, work that generic web-scraped text can’t fully substitute for as models push into more technical, professional domains. The company, founded three years ago by Brendan Foody, Adarsh H., and Surya Midha, reported 614 million dollars in gross revenue in the first half of 2026, up 70 percent from all of the previous year combined, with an annualized run rate exceeding 2 billion dollars by mid-year. Mercor pays out roughly 60 to 70 percent of that revenue directly to the contractors doing the actual labeling work, reflecting how labor-intensive high-quality data curation remains even at this scale.
Nvidia’s specific interest traces to Nemotron, its family of open-source models built to compete with other leading open-weight systems. A handful of Mercor staff now reportedly work almost entirely on the Nvidia account, and Nvidia paid the company tens of millions of dollars last quarter alone for expert-curated data supporting Nemotron’s development, even as Nvidia continues to lean heavily on synthetic, AI-generated data as well. Mercor’s broader client list includes OpenAI, Google DeepMind, and Anthropic, positioning it as critical infrastructure across several of the industry’s largest labs rather than a vendor tied to any single company.
A customer becoming an investor
General Catalyst, an existing Mercor backer, is reportedly leading discussions on the new round, though the precise size of Nvidia’s potential contribution and the round’s total size haven’t been disclosed, and talks remain preliminary with no deal finalized. This wouldn’t be Nvidia’s first move into the data-supply layer specifically: the company previously participated in rival data-labeling firm Scale AI’s 2024 funding round, which valued that company at 14 billion dollars, and Nvidia also sources training data from Turing separately. Taken together, that pattern shows Nvidia treating specialized data supply as strategic infrastructure worth direct financial backing, not just a category of vendor it pays and otherwise ignores.
The move fits inside a much larger pattern of Nvidia deploying capital well beyond its core chip business. The company invested 18.6 billion dollars into private companies and infrastructure funds in the quarter ending in April alone, already exceeding its total investment activity for the entire prior year, and it has separately mobilized more than 500 billion dollars in partnerships with financial institutions aimed at AI infrastructure. A Mercor stake, against that backdrop, is a comparatively small position, but a strategically pointed one: it deepens Nvidia’s control over a specific input, high-quality expert data, that increasingly determines how competitive an open-weight model like Nemotron can actually be against closed rivals.
Why the valuation jump is worth scrutinizing
Doubling a company’s valuation in under a year is an aggressive marker even in the current AI funding environment, and it’s worth separating Mercor’s genuine revenue growth from the multiple investors are willing to pay for it. Mercor’s revenue roughly matches the pace implied by its own reported figures, but a 20 billion dollar valuation on that base still implies a significant forward bet on continued acceleration, not simply a reflection of current performance. Whether that bet pays off depends heavily on whether demand for expert-curated training data keeps growing at its current pace as models mature, or whether synthetic data generation, which Nvidia itself already leans on substantially, gradually reduces how much labs need to pay for human-labeled data specifically.
Why chipmakers are moving into data supply at all
Nvidia investing in a data supplier is a genuinely different kind of move than the compute and infrastructure deals that dominate most AI investment headlines. Most of Nvidia’s capital deployment this year has gone toward securing demand for its own chips, backing cloud providers, model labs, and infrastructure buildouts that ultimately need GPUs. A stake in Mercor instead secures a different, less visible input further up the stack: the specialized human expertise that determines how good an open-weight model like Nemotron actually is once it ships. As frontier labs increasingly compete on post-training quality rather than raw pretraining scale alone, the value of high-quality, expert-labeled data has risen accordingly, and Nvidia backing one of its primary suppliers is a way of getting closer to, and arguably influencing, a resource it can’t simply manufacture the way it manufactures chips.
See PYMNTS’ full report for additional detail on the deal terms.
This kind of strategic backing sits alongside the broader trend covered in our explainer on how AI startup valuations are actually set.
